Evaluation of Terminology Translation in Instance-Based Neural MT Adaptation

Evaluation of Terminology Translation in Instance-Based Neural MT Adaptation M. Amin Farajian1,2, Nicola Bertoldi1, Matteo Negri1, Marco Turchi1, Marcello Federico1 1 Fondazione Bruno Kessler, Trento, Italy 2 University of Trento, Trento, Italy {farajian,bertoldi,negri,turchi,federico}@fbk.eu Abstract We address the issues arising when a neu- ral machine translation engine trained on generic data receives requests from a new domain that contains many specific tech- nical terms. Given training data of the new domain, we consider two alterna- tive methods to adapt the generic system: corpus-based and instance-based adapta- tion. While the first approach is compu- tationally more intensive in generating a domain-customized network, the latter op- erates more efficiently at translation time and can handle on-the-fly adaptation to multiple domains. Besides evaluating the generic and the adapted networks with conventional translation quality metrics, in this paper we focus on their ability to prop- erly handle domain-specific terms. We show that instance-based adaptation, by fine-tuning the model on-the-fly, is capable to significantly boost the accuracy of trans- lated terms, producing translations of qual- ity comparable to the expensive corpus- based method. 1 Introduction When deployed in production lines, machine trans- lation (MT) systems need to serve requests from various domains (e.g. legal, medical, finance, sports, etc.) with a variety of structural and lexi- cal differences. Considering that technical transla- tion (e.g. user guides, medical reports, etc.) rep- resents the largest share in the translation indus- try (Kingscott, 2002) and that a significant part c⃝ 2018 The authors. This article is licensed under a Creative Commons 3.0 licence, no derivative works, attribution, CC- BY-ND. of it deals with domain-specific terms, it is im- portant that machine translation delivers not only generic quality but also accurate translations of terms. The possibility of bearing different mean- ings in different contexts increases the difficulty of translating terms, making it an interesting and challenging topic in MT. Table 1 shows two ex- amples in which Google Translate1 (GT) and Bing translator2 (BT) wrongly translate domain termi- nology. In the first example, the English word ap- ple is wrongly recognized and translated as a term of the computer domain (apple) while it actually refers to the fruit type (mele). In the second ex- ample, on the contrary, Bing fails to recognize the multi-word term broken Windows by producing in- stead a literal translation that departs from the orig- inal sense. These examples show that existing MT systems still have difficulties in handling domain- specific terms, which calls for solutions to improve this aspect of MT. Ideal solutions for this real-world multi-domain translation scenario should be scalable enough to enable the industrial deployment at a reasonable cost, while guaranteeing a high level of flexibility in delivering good-quality translations for all (or most of) the domains. This is of higher impor- tance for the neural approach, where building the systems usually requires expensive GPU machines trained for several days to weeks on large amounts of parallel data. In this paper we analyze the ability of instance- based adaptation strategy in handling domain …

Content truncated to 500 words